Wire user-provided pinhole calibration into the standard input paths

[pb-evercoast]

Evidence, camera model, seeds, results and the reproduction package are in #105. This PR
is the calibrated-input wiring that issue asks for. Closes #56; related to #37 and #103.

pipeline.init.intrinsics=gt is exposed by the typed configuration, and
slam.optimize_intrinsics already resolves to false when it is selected. But neither
vipe infer video.mp4 nor vipe infer --image-dir has a way to hand a calibration to the
pipeline: RawMp4Stream and FrameDirStream never populate frame intrinsics, and
DefaultAnnotationPipeline unconditionally appends GeoCalib and asserts that no intrinsics
are present. Selecting gt from the CLI therefore fails for every user who already knows
their camera — which is the common case for rendered, rig-calibrated or benchmark footage
(#37, #56).

This PR completes both paths with no change to the default behaviour:

• vipe infer <video.mp4 | VIDEO_DIR | --image-dir DIR> --intrinsics calibration.json
  accepts one pinhole calibration ({"width", "height", "fx", "fy", "cx", "cy"}), validates
  it before any model is loaded, checks it against the frame size of every input it is
  applied to (a directory of videos shares one camera; each video is checked in turn and a
  mismatch names the video), and selects pipeline.init.intrinsics=gt (which disables
  intrinsic optimization through the existing ${neq:...} resolver).
• RawMp4Stream and FrameDirStream take an optional [fx, fy, cx, cy] tensor, validate
  it (shape, finiteness, positive focal lengths), stamp a per-frame copy plus
  CameraType.PINHOLE on every VideoFrame, and advertise INTRINSICS and CAMERA_TYPE
  in attributes() so the pipeline's attribute check sees them through the cache wrapper.
• DefaultAnnotationPipeline._add_init_processors branches on init.intrinsics:
  geocalib keeps the current behaviour and assertions unchanged; gt skips GeoCalib and
  refuses a stream that does not carry intrinsics and a camera type, rather than silently
  estimating them; any other value raises. PoseOnlyAnnotationPipeline and
  PoseOnlyLongAnnotationPipeline inherit this method, so the calibrated path is available
  to them without a change of their own.
  Nothing changes for an invocation without --intrinsics: no new flag is required, no new
  file is written, and the RNG state is untouched.

The --seed flag and the resolved-config.json record that the ablation in #105 relied on are
deliberately not in this PR. They are a separate concern (a per-run record of the resolved
configuration, not a determinism claim) and can follow as their own change if wanted.

Tests

tests/test_calibrated_frame_dir_input.py (pytest, same layout as the existing suite):

• streams, both FrameDirStream and RawMp4Stream: uncalibrated advertises nothing;
  calibrated advertises exactly {INTRINSICS, CAMERA_TYPE} and the advertisement survives
  ProcessedVideoStream; malformed / non-finite / non-positive-focal intrinsics are refused
  with the documented messages; every frame carries its own copy of the intrinsics on CUDA
  (GPU-only);
• config: init.intrinsics=gt resolves slam.optimize_intrinsics=false, geocalib
  keeps it true;
• pipeline: gt refuses a stream without intrinsics, adds no GeoCalib processor for a
  calibrated stream, and the geocalib branch still asserts against pre-existing
  intrinsics;
• CLI: --intrinsics is accepted on the video path as on the frame directory; malformed
  calibration JSON is rejected before any model loads and before the output directory
  exists; a calibration whose dimensions disagree with the frames is rejected (GPU-only);
  a directory of videos is checked video by video and the refusal names the offending file
  (GPU-only).

CUDA-dependent cases are skipif-guarded; the rest run on CPU.

Compatibility

Pure Python; no changes to vipe_ext or the compiled extensions. The v1.2.0 form was
exercised on an A100 against torch 2.13.0+cu130; the upstream CI matrix (cu124 /
cu128) uses only the same basic tensor API. The rebased form keeps upstream's new
behaviour intact: the eager .cache() is still skipped for the long-sequence pipeline, and
the per-video loop is unchanged apart from the calibration check.

What this PR does not do

• It does not add a --seed flag or write a record of the resolved configuration; see above.
• It does not add distortion coefficients to the calibration file. VideoFrame.intrinsics
  already documents a (4+D,) layout; a follow-up could accept k1.. for the camera
  models that consume them. This PR is pinhole-only on purpose.
• It does not change any default or config file. Without --intrinsics every path behaves
  exactly as before.

Tests on hardware

On an A100 (sm_80), from a clean clone at main 8c9f361 with this commit applied by git am and vipe_ext rebuilt: tests/test_calibrated_frame_dir_input.py 28 passed, and the whole tests/ suite on the same tree 108 passed. CUDA-dependent cases are skipif-guarded for CPU CI.